[Alex McGregor]

Hello Carlo,

My apologies for the late reply. I wanted to let you know that I have run the regression and got significant results.

However, my regression gives me a very low R-Squared (0.0002-0.0004). I have read that a low r-squared can be expected for stock returns as well as that it does not necessarily mean my model is not good, especially because I am not trying to predict something and also got significant results. Or does this low r-squared definitely mean I will have to change my model by adding more variables (omitted variable bias) such as company beta or other independent variables?

[Carlo Lazzaro]

Alex:
R-sq magnitude is also research field-related and every general advice on this issue can be misleading.
Hence, if other publications covering similar topics report a low _R_sq you should not worry.

[Alex McGregor]

Thank you Carlo.

I have been looking at publications related to quantitative easing and stock returns, and the R squared reported are never that low. Does that mean I would have to add more variables or can I still argue that I am not trying to predict the dependent variable as I have statistically significant results.

[Alex McGregor]

Hello Carlo,

Regarding the low R-squared, I have one model measuring stock returns (R square of 0.06) vs model measuring ABNORMAL stock returns (0.005) with both coefficients at approximately the same p-value. Would it be more appropriate to chose the normal stock return model due to the higher R-square?

[Carlo Lazzaro]

Alex:
R-sq magnitude and/or statistical significance of coefficients are not the right standards to choose among different regression specifcations.
Following the literature in your research field, you should run the appropriate regression with all the predictors that give the fairests and truest view of the data generating process that you're investigating.

[Alex McGregor]

I do not understand why they are not the right standards. Should I instead be looking at BIC / AIC or what are the standards to choose among different regression specifications?

[Carlo Lazzaro]

Alex:
if you have to choose between, say, two OLS with the same regressand but different predictors, you should look at the adjusted R-sq (not R-sq).
In addition, you can obtain statistical significant (but misleading) coefficients from an ill-specified regression.
This is the reason why regression specification should be your first concern.

[Ramshah Lone]

Hi,
I am writing in this forum as i think my question is related to the discussion going here. I am doing a study in which i am interested in looking at impact of Male CEO on firm risk taking. The data is from 2004-2018 and there are 120 firms. These 120 firms have been selected from top 500 firms whose annual reports and stock data was available. I need some guidance that when i run FE (as also supported by Hausman test) can I set my data as

1. xtset Year Firm
xtreg dep var indp var1 indp var2..., fe

2. My second question is that do i need to put vce (cluster/robust) at the end of FE or stata can adjust it

3. Should the answer of my regression be same if i write xtset Frim Year

4. How do you check mediation effect in stata ( i know how to do moderation)

Thanks
Ramshah

[Carlo Lazzaro]

Ramshah:
welcome to this forum.
As far as your questions are concerned:
1) Your -xtset- code is wrong.
The correct one is:

xtset Firm Year

.
The sequence of -xtset- terms should not be reversed.
You regression code is correct (but delete the blanks, such as -dep var-, as Stata considers them as mistaken chunks of code):
2) yes, you should, as you have 120 panels (ie, clusters) in your dataset. I would also add, among the predictors, -i.Year-;
3) see reply to 1);
4) see: https://www.stata.com/meeting/italy1...t13_grotta.pdf.

As an aside, please read (and act on) the FAQ about how to post more effcìectively (and increase, in turn, your chances of getting more helpful replies). Thanks.